2016/04/28
Introduction
How do we move towards a more data driven decision making for conservation & management?
We have lots of data
We have lots of models
So what could go wrong?
Outline
Part I: Making the data fit the models
- Model here, data there
- Long tail data
- Scaling
Part II: Making the model fit the data
- Decision Theory
- POMDPs
Part I: Ecoinformatics – making the data fit the models
Expanding sources of data
Remote Sensors
Microsensors
NEON
Ocean Observatory Initiative
Data publication
Model here, data there
Data access
- Online databases written for humans but not computers
- Online databases written for computers but not humans
- Written for the wrong humans
rOpenSci
rOpenSci
rOpenSci
Not written for computers
Not written for humans
Data written for the wrong humans?
APIs
rOpenSci Packages
rOpenSci data packages aim to be:
Written for humans
- Provide a simple and logical query language for accessing data
Written for computers
- Provide a programmatic interface to the data: It should be possible to automate repetitive data access tasks.
Format friendly
- Provide the data in formats where modeling can be most readily applied (
data.framesinstead of XML, format exchange methods)
Long tail data
Beyond databases: the long tail of academic data
Long tail data
Vertically Integrated Repositories =
Metadata Repository = Data Lake
Metadata Repository
rOpenSci data publication tools aim to:
- (semi)automate metadata capture
- Benefit the data creator first
- Reflect best practices
- Leverage advances in informatics & data science
https://github.com/ropensci/EML https://github.com/ropensci/RNeXML https://github.com/ropensci/datapack
Workflow
Data Synthesis
Semantics
Long tail data
- Data discovery
- Data management
- Data synthesis
- Data publication
Scaling
Data doesn't fit the computer
- Most computation is still local
- Scaling beyond the laptop usually involves complete re-tooling (and to far less friendly tools)
- Doesn't have to be this way
Rocker
- Portable environments: from laptop to HPC
- Reproducible environments
- Community-optimized environments
Jetstream
Part II: Optimal control – making the model fit the data
Methods ignore data realities
- missing data
- non-random bias
- data heterogeneity
- non-stationarity
- imperfect observations
Today, we'll focus on just one:
- imperfect observations
Consider optimal control problems:
- Practical application, clear objectives
- Decision theory: optimal control of stochastic systems …
- … but typically still requires strong assumptions
Why Optimal Control?
- Cannot have an answer without a question
- Distinguish between what we need to know and what doesn't matter
- Consider "actions" as well as "states"
Sumatran Tiger
Sumatran Tiger
Spartina
Lambert et al (2014) Science
Spartina
Fisheries
Fisheries
Markov Decision Processes
Partially Observed Markov Decision Processes
Decision Matrix
Decision Matrix
Decision Matrix
Decision Matrix
Decision Matrix
State equation
- If we want to make decisions about the future, we need a predictive model
\[ \vec X_{t+1} = f(\vec X_t, \vec a_t, \xi) \]
Decision Matrix
POMDP Algorithms
- Monahan's Enumeration Algorithm (1971, 1982)
- Sondik's One-Pass Algorithm (1971)
- Cheng's Linear Support Algorithm (1988)
- Littman, et al's Witness Algorithm (1994)
- Zhang's Incremental Pruning Algorithm, (1996)
Modern POMDP Algorithms
- PBVI (Pineau et al 2003)
- Perseus (Spaan, & Vlassis 2005)
- HSVI (Smith, & Simmons 2005)
- GapMin (Poupart et al 2011)
- PEMA (Pineau & Gordon 2005)
- FSVI (Shani et al 2008)
- SARSOP (Kurniawati et al 2008)
Does partial observability matter?
Classic, fully-observed solution:
POMDP solutions
Impact on management
Acknowledgements
Collaborators:
- Milad Memarzadeh
- Karthik Ram
- Scott Chamberlain
- Matt Jones
Funding:
- NSF, Helmsley
- ESPM